mlsanityData to 14 Oct 2026

Compositional Reasoning Depth Predicts Clinical AI Failure: Empirical Evidence Consistent with Transformer Compositionality Limits in Electronic Health Record Question Answering

arXiv 2606.16890 ↗First posted 15 Jun 2026cs.CL cs.AIcited by 0

On the map: LLM reasoning → Clinical and biomedical applications

Abstract (arXiv)

Aggregate accuracy benchmarks conceal a systematic structure in how large language models fail at electronic health record (EHR) question answering: questions requiring more inferential steps produce disproportionately more errors. Motivated by theoretical results on transformer compositionality limits, we introduce a pre-specified hop-count taxonomy -- the number of distinct reasoning steps required to answer a clinical question from an EHR -- as a principled predictor of model failure. We annotate 313 clinician-generated MedAlign EHR question-answer pairs across four hop levels and evaluate 301 questions in a within-model ablation (claude-sonnet-4-6, zero-shot vs. extended thinking) and cross-architecture replications (gpt-4o and gpt-5.4-2026-03-05, zero-shot). All three models, spanning two providers and two OpenAI generations (GPT-4 and GPT-5), show monotone accuracy decline with hop count: Claude Sonnet zero-shot falls from 30.6% (hop=1) to 17.6% (hop=4) (Cochran-Armitage z=-2.30, p=0.011; OR per hop 0.72, 95% CI [0.56,0.92], p=0.008); GPT-4o replicates this (37.8% to 14.7%; OR 0.58 [0.45,0.75], p<0.001); and gpt-5.4-2026-03-05 confirms it (37.8% to 23.5%; OR 0.80 [0.66,0.98], p=0.027). A pre-specified context-sufficiency audit shows higher-hop questions are not differentially disadvantaged by EHR truncation (answerability 93-95% at hops 2-4 vs. 79% at hop=1), so the decline reflects compositional reasoning difficulty. Extended thinking did not significantly flatten the accuracy-depth curve across three reasoning conditions, and thinking-token usage scaled with hop count (r=0.31, p<0.0001), consistent with the predicted O(k) computational requirement. Hop count is thus a theory-motivated, cross-architecture predictor of large-language-model error on EHR question answering, with direct implications for deployment risk stratification of clinical AI.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Clinical and biomedical applications, one of 17 directions in LLM reasoning. Builds on 16 of its 28 references that are in the map. First posted on arXiv 15 Jun 2026.

Builds on

16 of 28 references.

Closest ancestors · 1

Same direction, newest first

  1. 2026FDARxBench: Benchmarking Regulatory and Clinical Reasoning on FDA Generic Drug Assessment20 Mar

Other references · 15

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  2. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  3. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  4. 2022Large Language Models are Zero-Shot Reasonerscited by 4,374
  5. 2018HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answeringcited by 3,754
  6. 2020What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Examscited by 1,356
  7. 2020Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Stepscited by 1,282
  8. 2024OpenAI o1 System Cardcited by 1,244
  9. 2021MuSiQue: Multihop Questions via Single-hop Question Compositioncited by 916
  10. 2021Show Your Work: Scratchpads for Intermediate Computation with Language Modelscited by 886
  11. 2024On Limitations of the Transformer Architecturecited by 63
  12. 2023MedAlign: A Clinician-Generated Dataset for Instruction Following with Electronic Medical Recordscited by 56
3 more
  1. 2023Brainformers: Trading Simplicity for Efficiencycited by 31
  2. 2025Compositional Reasoning with Transformers, RNNs, and Chain of Thoughtcited by 6
  3. 2025First, do NOHARM: a medical safety benchmark and randomized study of physician and AI teaming on clinical consultationscited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026ART: Action-based Reasoning Task Benchmarking for Medical AI Agents
  2. 2025Limitations of Large Language Models in Clinical Problem-Solving Arising from Inflexible Reasoning
  3. 2026UIC-AIHealth4All at ArchEHR-QA 2026: Answer-First Evidence Grounding for Clinical Question Answering
  4. 2026Yale-DM-Lab at ArchEHR-QA 2026: Deterministic Grounding and Multi-Pass Evidence Alignment for EHR Question Answering
  5. 2025FHIR-AgentBench: Benchmarking LLM Agents for Realistic Interoperable EHR Question Answering
  6. 2022Measuring and Narrowing the Compositionality Gap in Language Models
  7. 2026Advances in LLM Reasoning Enable Flexibility in Clinical Problem-Solving
  8. 2024HealthQ: Unveiling Questioning Capabilities of LLM Chains in Healthcare Conversations